WHAT PREDICTIVE ACCURACY IS REQUIRED FOR AN ARTIFICIAL INTELLIGENCE TOOL TO BE COST-EFFECTIVE IN PREDICTING DIALYSIS NEED AMONG PATIENTS WITH CHRONIC KIDNEY DISEASE?
Author(s)
Lisa Masucci, MSc1, Martin Klamrowski, MASc1, Ran Klein, PhD2, Christopher Mccudden, PhD1, James Green, PhD3, Amit Garg, MD4, Babak Rashidi, MD1, Ayub Akbari, MD1, Deena Fremont, MSc2, Gregory Hundemer, MD1, Kednapa Thavorn, PhD5.
1The Ottawa Hospital, Ottawa, ON, Canada, 2University of Ottawa, Ottawa, ON, Canada, 3Carlton University, Ottawa, ON, Canada, 4London Health Sciences Centre, Ottawa, ON, Canada, 5Ottawa Hospital Research Institute, Ottawa, ON, Canada.
1The Ottawa Hospital, Ottawa, ON, Canada, 2University of Ottawa, Ottawa, ON, Canada, 3Carlton University, Ottawa, ON, Canada, 4London Health Sciences Centre, Ottawa, ON, Canada, 5Ottawa Hospital Research Institute, Ottawa, ON, Canada.
OBJECTIVES: Chronic Kidney Disease (CKD) impacts 11 to 13% of adults, and 50% of patients progressing to kidney failure initiate dialysis unexpectedly in hospital. Unplanned dialysis leads to worse clinical outcomes, lower quality of life, and substantially higher healthcare costs. An artificial intelligence (AI) tool has been developed to predict dialysis need at 6 and 12 months using routinely collected CKD clinic data. This study determined the minimum predictive accuracy required for the tool to be cost-effective.
METHODS: A state-transition Markov model was developed to estimate lifetime costs (2025 Canadian dollars) and quality-adjusted life-years (QALYs) for the AI tool compared with usual care from the Canadian public healthcare payer perspective. The model simulated adults with CKD transitioning among CKD without dialysis, pre-emptive kidney transplant, hemodialysis, peritoneal dialysis, unplanned dialysis, and kidney transplant health states over 8 years. AI tool accuracy, costs, and utilities were obtained from the published literature. Time-varying transition probabilities between health states were informed by real-world administrative data. Costs and QALYs were discounted by 1.5% annually. Threshold and probabilistic analyses were conducted.
RESULTS: At a sensitivity of 0.651 and specificity of 0.97, the AI tool resulted in 0.00055 additional QALYs and $26.95 per patient compared to usual care, resulting in an incremental cost-effectiveness ratio (ICER) of $51,964/QALY gained. Threshold analyses demonstrated substantial sensitivity of economic outcomes to predictive performance. Reducing sensitivity from 0.651 to 0.645 increased the ICER to CAD $187,305/QALY gained. Results were primarily driven by reductions in unplanned dialysis initiation and downstream costs and outcomes.
CONCLUSIONS: The economic value of AI-enabled dialysis prediction depends on maintaining high predictive performance, highlighting the importance of rigorous validation and ongoing monitoring of AI tools before large-scale implementation. These findings provide decision-makers with a quantitative performance threshold to inform adoption, reimbursement, and evaluation of AI technologies in CKD care.
METHODS: A state-transition Markov model was developed to estimate lifetime costs (2025 Canadian dollars) and quality-adjusted life-years (QALYs) for the AI tool compared with usual care from the Canadian public healthcare payer perspective. The model simulated adults with CKD transitioning among CKD without dialysis, pre-emptive kidney transplant, hemodialysis, peritoneal dialysis, unplanned dialysis, and kidney transplant health states over 8 years. AI tool accuracy, costs, and utilities were obtained from the published literature. Time-varying transition probabilities between health states were informed by real-world administrative data. Costs and QALYs were discounted by 1.5% annually. Threshold and probabilistic analyses were conducted.
RESULTS: At a sensitivity of 0.651 and specificity of 0.97, the AI tool resulted in 0.00055 additional QALYs and $26.95 per patient compared to usual care, resulting in an incremental cost-effectiveness ratio (ICER) of $51,964/QALY gained. Threshold analyses demonstrated substantial sensitivity of economic outcomes to predictive performance. Reducing sensitivity from 0.651 to 0.645 increased the ICER to CAD $187,305/QALY gained. Results were primarily driven by reductions in unplanned dialysis initiation and downstream costs and outcomes.
CONCLUSIONS: The economic value of AI-enabled dialysis prediction depends on maintaining high predictive performance, highlighting the importance of rigorous validation and ongoing monitoring of AI tools before large-scale implementation. These findings provide decision-makers with a quantitative performance threshold to inform adoption, reimbursement, and evaluation of AI technologies in CKD care.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
P24
Topic
Economic Evaluation
Disease
Urinary/Kidney Disorders